Prediction of rock mechanical properties from geochemical signatures using machine learning algorithm

This paper presents a data-driven model built using machine learning technique, namely Support-Vector Regression (SVR), for predicting rock mechanical properties (RMP) of tight sands and shale formations based on measured elemental information. This study utilized 324 data points obtained from Paleozoic sequences covering 8000 ft interval. Elemental data was acquired by scanning intact samples on MicroXRF machine, whilewas used for modelling as well as for establishing lithofacies. RMP were obtained at atmospheric conditions using a non-destructive impulse hammer technique. SVR was used to establish a correlation between elemental geochemical information with RMP. The dataset was categorized into two: training and testing dataset using a 70-30 percent split. Training data was used to train SVR and establish the RMP prediction model by tuning the hyperparameters. Testing data was used to evaluate the predictive model by comparing the predictions with the actual measurement. Two quantitative measures for estimation accuracy including mean absolute error and cross-correlation plots between the actual and estimated RMP were employed to evaluate the prediction accuracy. The results demonstrate that the estimated RMP has a good agreement with the actual RMP, which is indicated by the small error and high coefficient of determination. Moreover, we used Adjacent Boosting technique to perform feature selections and the results show that the SVR-based model still generates good prediction when the number of features in the input is reduced by more than half. This study demonstrates that a reliable predictive model can be built with few intrisinc features, in the absence of robust mineralogical and elemental information.